Collin Phillips, Virginia Tech; Aster Saul, Virginia Tech; Kevin Schroeder, Virginia Tech; Jonathan Kadan, SSC/SYD 85/BI – Odyssey
https://doi.org/10.64861/CKHP9224
Keywords: Space Domain Awareness, Autonomous Sensor Tasking, Information Gain, Serendipitous Observation
Abstract:
Current tasking systems typically assign a single sensor to a single expected RSO by way of bipartite graph matching, ignoring the potential for serendipitous observations of other targets in the field of regard. These serendipitous observations allow sensors to collect more information during a tasking period than was otherwise expected by gathering unintended observations of nearby RSOs. Previous work has directly exploited this serendipity by solving for tasks using sensor boresight vectors against sets of multiple RSOs for increased information gain. However, this approach still leaves the problem of relative motion unaddressed, instead abstracting away some of the intricacies of actual collection schedules. In this study, we aim to improve upon that strategy by incorporating the relative motion of each task-relevant RSO to develop a motion plan for boresight tasking.
Leveraging recent work in optimizing for serendipity, namely “Luck is a Skill”, we develop improvements with a set of extensions which incorporate boresight motion and detection capabilities of the post-processing pipeline. To achieve this objective, we incorporate a rate-based observation model, and validate its performance against the detection capacity of a relevant post-processing pipeline by simulating high-fidelity observations offline. This provides a reference from which we augment the detection model in our existing simulation and tasking software. Using these models, we develop our algorithm to generate information optimized boresight-rate tasking schedules. We call our complete extended approach LIASSON. The LIASSON algorithm inherits a hierarchical planning system which first recursively solves for a set of valid local sensor actions, then deconflicts them into a joint global task with coordinate ascent. There are two ways to solve for the local tasks: scan line or convolutional. The convolution approach first discretizes the field of regard into a grid and pools particles sampled from the estimators for each visible member of the catalog into their respective cells. Then a filter corresponding to the sensor’s field of view is convolved across the grid to generate an approximate information map.
The first of our proposed modifications includes additional anisotropic “motion” filters along additional channels of the convolution. By first tracing estimated RSO paths along the grid, convolving each motion filter approximates information gain along the corresponding rate-track. The output of this step provides a set of pointing directions and motion vectors for a collection. By taking a subset or linear combination across these channels, we can construct an approximate information maximizing task. In both versions of our algorithm, scan line and convolutional, the tasking process is aware of the probability of observing every RSO in its field of regard during each potential boresight operation. Our second approach simply produces a rate track of the exponential weighted average of those RSOs where the rate corresponds to information gain. The final extension simply takes the top-k subjects and generates a multi-rate-track schedule for collecting for each of them.
We conduct an analysis on the theoretical benefits of each approach, the implementation details of both the reduced-order observation model and our LIASSON tasking algorithm, as well as a comparison approach to prior work. We evaluate the catalog covariance over time, average observation age, and quality of the observations gathered. Additionally, we compare the motion-aware version of our simulation with an instantaneous model to evaluate the impact of rates on simulation fidelity and observation quality. Our approach demonstrates an improvement over previous methods by incorporating rate track scheduling into a state-of-the-art approach to information optimizing task scheduling.
Date of Conference: September 15-18, 2026
Track: Space Domain Awareness